Does the Gettier Problem Show That Knowledge Cannot Be Analysed?

Bu yazıyı site için hazırladım, bir ders ödevi değil. Bölümdeki metinlerin nereden geldiğini girişte ayrı ayrı yazdım.

---

Does the Gettier Problem Show That Knowledge Cannot Be Analysed?

1. The Question Behind the Question

Gettier's (1963) three-page paper is usually described as having refuted the justified-true-belief analysis of knowledge. That description is accurate but uninformative, because it leaves open the interesting question. Did Gettier show that this analysis was wrong, or did he show something structural — that no analysis of this kind can work?

The two readings recommend very different research programmes. On the first, epistemology after 1963 is a repair job: find the missing fourth condition. On the second, the repair job is hopeless in principle, and the right response is to stop analysing. Williamson (2000) takes the second route and treats knowledge as unanalysable and explanatorily primary. Most of the field, for several decades, took the first.

I will argue that the strongest anti-analysis argument — Zagzebski's (1994) — does not quite establish unanalysability, but that it establishes something almost as damaging: any analysis that adds conditions to justified true belief will either be Gettierisable or will smuggle in a factivity that makes the analysis uninformative. The conclusion I reach is that knowledge may well be analysable, but not reductively; and that this is a substantive result rather than a defeat.

2. The Structure of a Gettier Case

It helps to state the recipe rather than the examples, because the recipe is what generalises.

Take a belief that is justified but false. Derive from it, by a valid inference that preserves justification, a belief that happens to be true. The derived belief is justified (justification transmits across competent deduction), true (by stipulation), and believed. Yet we do not want to call it knowledge, because its truth has nothing to do with the justification: the belief is true by accident relative to the grounds on which it is held.

Zagzebski (1994) makes the recipe explicit. Suppose a proposed analysis says that knowledge is true belief plus some further condition C, where C is not itself factive — that is, where it is possible to satisfy C and still have a false belief. Then, she argues, we can always construct a case as follows: take a belief satisfying C that is false, then modify the scenario so that the belief becomes true by some accident unconnected to C. The result satisfies the analysis and is not knowledge.

The argument has the form of a dilemma. Either C is factive, or it is not. If it is not, the recipe applies. If it is, the analysis contains a condition that already entails truth, and the question arises whether the analysis is informative or merely restates that knowledge entails truth.

3. Repairs and Their Costs

It is worth seeing how the major repairs fall under the dilemma, because the pattern is more instructive than any single failure.

No false lemmas. Require that the justification not essentially depend on any false belief. This handles the classic cases elegantly. But Goldman's (1976) fake barn case does not involve any false lemma: Henry looks at a real barn, believes there is a barn, and his perceptual belief is not inferred from anything false. The condition is too narrow.

Causal theory. Goldman (1967) requires an appropriate causal connection between the fact and the belief. This handles fake barns poorly too, and it struggles with knowledge of universal generalisations and of mathematical truths, where the relevant fact is not a causal relatum at all.

Defeasibility. Lehrer and Paxson (1969) require that there be no true proposition which, if added to the subject's evidence, would defeat the justification. The difficulty is well known: misleading defeaters. There are truths whose addition would defeat a justification that we intuitively regard as knowledge-conferring, and excluding them requires a distinction between genuine and misleading defeaters that is hard to state without appealing to knowledge.

Sensitivity. Nozick (1981) requires that the subject would not believe p if p were false. The condition delivers verdicts on many Gettier cases but has the notorious consequence that knowledge is not closed under known entailment, which most epistemologists treat as too high a price.

Safety. Pritchard (2005) and Sosa (2007) require that the belief could not easily have been false. This handles the cases well and does not obviously break closure. But safety is a modal condition on the truth of the belief, and here Zagzebski's dilemma bites: safety is close to factive in the relevant sense, since it constrains truth across nearby worlds. Whether it is informative depends on whether we have an account of "nearby" that does not itself presuppose epistemic notions.

The pattern is striking. Each repair handles the cases that motivated it and is defeated by a case designed against it. Weatherson (2003) argues that this pattern should make us suspicious of the method rather than of the analyses: if a theory is otherwise systematic and simple, a counterexample that costs us a great deal to accommodate may be a reason to revise our intuitions instead.

4. Why the Anti-Analysis Conclusion Does Not Follow

Williamson (2000) draws the strong conclusion: knowledge is a primitive mental state, not a conjunction of belief, truth and something else. The Gettier literature's failure is then explained rather than merely recorded.

I think the argument from failure to unanalysability is weaker than it looks, for three reasons.

First, the inference from "every proposed analysis has failed" to "no analysis can succeed" is inductive and the sample is not large. Analyses of other philosophically loaded concepts — causation, for instance — have similarly long failure records without producing a consensus that they are primitive.

Second, Zagzebski's dilemma, read carefully, does not conclude that analysis is impossible. It concludes that any non-factive additional condition is Gettierisable. That leaves open the option of a factive condition that is nevertheless not trivial. Sosa's (2007) apt belief — belief true because of competence — is a candidate: the "because" relation is factive but the analysis is not thereby empty, since the explanatory relation does substantive work.

Third, the demand that an analysis be reductive is a demand we do not impose elsewhere. We can offer an illuminating analysis of a notion in terms that include cognate notions, provided the account explains something. If "knowledge is apt belief" tells us what unifies the cases, it is not disqualified merely by failing to reduce knowledge to non-epistemic materials.

5. What the Dilemma Does Establish

Granting all that, the dilemma establishes something worth having, and I think this is the correct verdict rather than either of the standard ones.

If Zagzebski is right, then the classical project — analysing knowledge into independently specifiable, non-factive components — is finished. Not merely unsuccessful so far, but structurally unable to succeed, because the recipe for generating counterexamples is general.

That is a significant result, and it reframes what a successful analysis would have to look like. It would have to include a condition that links the truth of the belief to the grounds of the belief in a way that is not accidental: a "because" relation, an explanatory connection, a manifestation of competence. And any such condition will be factive in the sense that satisfying it guarantees truth.

The remaining question is whether such an account is informative. I think it can be, on the following test: an analysis is informative if it explains the pattern of our verdicts across cases, including cases we had not considered when constructing it. Virtue-theoretic accounts have some claim to this. They explain fake barns as cases where the belief is true but not because of competence, since the competence would have produced a false belief in the immediate environment. They explain the classical cases in the same way. And they do not need to be repaired case by case.

6. A Complication About Method

There is a methodological worry that cuts across the whole debate and that I do not think can be dismissed.

The entire Gettier literature runs on intuitions about cases. If those intuitions are unstable across populations, or sensitive to the order in which cases are presented, then the data being fitted are less solid than the theories built on them. The experimental literature on philosophical intuitions has raised this concern seriously enough that it cannot simply be waved away, though the picture is contested and later studies have found more convergence on Gettier verdicts than early results suggested.

This does not dissolve the problem, but it does affect what "conclusive" can mean here. The claim that the classical analysis fails does not rest on a single fragile verdict; it rests on a recipe that produces the same verdict across indefinitely many cases, and the robustness of that pattern is better evidence than any one case.

7. Conclusion

Gettier did not show that knowledge cannot be analysed. He showed, and Zagzebski made explicit, that knowledge cannot be analysed as true belief plus a non-factive condition, because such analyses are generatively vulnerable to a single recipe.

The honest conclusion is therefore narrower than "knowledge is unanalysable" and broader than "we have not yet found the fourth condition". It is this: the accidental connection between grounds and truth is not an additional defect to be excluded by a further clause, but the thing an account of knowledge has to be built around. Analyses that start from that relation are still in the running. Analyses that treat it as a residue to be patched are not.